Research On Several Key Technology Of Disease Analysis In SD-OCT Retinal Images

Visual impairment is an important factor affecting people's quality of life,while retinopathy is a major cause of vision impairment,such as age-related macular disease(AMD),glaucoma,diabetic retinopathy(DR).In recent years,optical coherence tomography(OCT)imaging technique has been widely used in the diagnosis of retinal diseases,especially in the frequency domain OCT(SD-OCT)imaging technology.Due to its greater imaging speed,high-resolution,and non-invasive,retinal layers can be identified in SD-OCT images,and high resolution cross-section images can be provided for clinicans.Image processing and pattern recognition techniques to quantitatively analyze retinal diseases in SD-OCT images are becoming imprortant in the auxiliary diagnosis and treatment of retinal diseases,which has significant clinical and research value and also has great challenges.On one hand,with the extensive application of SD-OCT imaging technology,efficient segmentation and quantification of the retinal diseases is becoming more and more important.On the other hand,SD-OCT imaging technology provides more clinical phenotypes related to retinal diseases.However,developing an algorithm to predict the future disease progression and evaluate the efficacy of the treatment is a hot research topic in the field of Ophthalmology.Therefore,the thesis focuses on retinal layers segmentation,macular fovea center identification,segmentation and quantification of GA region and GA progression prediction.The major contributors are as follows:(1)Two retinal layers segmentation algorithms in SD-OCT images are proposed,which are automated layers segmentation algorithm using dual gradient and spatial correlation smoothness constraint,and layers segmentation method based on multi-scale 3-D graph search method.For the normal eyes,these two algorithms accurately segment retinal layer boundaries.For abnormal images,the first segmentation algorithm has a limited ability to segment the retinal layers,while the second algorithm can automatically detect retinal layer boundaries in SD-OCT images with AMD,and its segmentation results are relatively accurate.Experimental results demonstrated that the proposed method has better accuracy than other segmentation methods.(2)An automatic algorithm to detect macula foveal center in SD-OCT images is proposed.The pathological characteristic of the foveal center in SD-OCT images is the lowest point in the foveal depression,and this characteristic is normally taken as the key feature in fovea center determination.An improved saliency method with local orientation features,local data structure features and location information is developed to identify the macular fovea region.The foveal center is identified by searching the location of minimum total retina thickness within macular fovea region.Quantitative experimental results demonstrated that the proposed algorithm can achieve good performance for normal and abnormal cases with different severe disruption of the retinal anatomy.(3)An improved active contour model via local similarity factor is proposed,which is applied to the segmentation and quantification of GA lesions.The spatial distance in the model is included to balance the intensity differences between neighboring pixels and global mean intensity on two sides of active contour.Then an initial coarse GA region is estimated by an iterative threshold segmentation method and an intensity profile set,and subsequently refined by the improved model.Quantitative experimental results proved that the proposed model shows good GA segmentation results.(4)Being able to predict the locations of future GA involvement could be important for emerging therapies,it remains a challenging problem that is unsolved to date.19 comprehensive quantitative imaging features as predictors of future GA growth are extracted from SD-OCT images.We develop and evaluate a model to predict the magnitude and location of GA growth at given future times using the quantitative features in three possible scenarios,where the prediction model is built by a random forest classifier.Experimental results demonstrated the potential ability of our predictive model to predict future regions where GA is likely to grow and to identify the most discriminant early indicator of regions susceptible of GA growth.